Provider Comparison

Claude Opus 5 vs. Claude Fable 5 (2026): Half the Price, Half a Percent Off the Peak

Claude Opus 5 vs Claude Fable 5 in 2026: list price, cost per completed task, knowledge cutoff, caching and where Fable still holds the peak.

5
The new default
vs
2
The frontier ceiling
Quick Verdict

Opus 5 is the right default for almost everyone, and the argument is not that it is the smarter model — it is that it removes the reason to pay double. Half the list price, half the cache-hit price, a lower cache minimum, a four-month fresher knowledge cutoff and a faster latency rating are not marginal gains that need a benchmark to justify them; they show up on the invoice and in agent runtime whether or not the leaderboard moves. On Anthropic's own task-level numbers the cost story is stronger still: half the cost per task on CursorBench 3.2 for a score within 0.5% of Fable 5's peak, and a third of the cost on OSWorld 2.0 for a better result. What Fable 5 keeps is narrow but real. It still holds the peak on the benchmark Anthropic used to make the cost argument, the Mythos class remains ahead on cybersecurity evaluations, and it has a public record where Opus 5 has a launch post. That last point deserves more weight than it usually gets: at the time of writing, the numbers that make Opus 5 look dominant are Anthropic's own charts, published two days ago, and the one third-party leaderboard we checked labels its #1 placement "Estimated" while Fable 5's lower placement is "Supported". That is not evidence that the claims are wrong — it is a reason to verify them on your own workload before you re-platform a production pipeline. The practical decision splits cleanly. If you are on Opus 4.8, move: it is the same price and a strictly better model by every published measure, so the only work is re-checking max_tokens now that thinking is on by default. If you are on Fable 5, run one honest evaluation before switching — if Opus 5 holds on your tasks, you halve a recurring bill; if your workload lives in the last half percent, you have just paid for the evidence that Fable 5 is worth its price. And if you were about to pay Fable rates purely for speed, note that Opus 5's Fast mode costs exactly the same $10 / $50, which makes that a direct comparison rather than a tier decision.

Detailed Comparison

A side-by-side analysis of key factors to help you make the right choice.

Factor
The new defaultRecommended
The frontier ceilingWinner
List API price
$5 per million input tokens, $25 per million output tokens.
$10 per million input tokens, $50 per million output tokens — exactly double.
Cost per completed task
On CursorBench 3.2 at max effort it lands within 0.5% of Fable 5's peak score at half the cost per task; on OSWorld 2.0 it beats Fable 5's best result at just over a third of the cost.
Holds the peak score on those same benchmarks, but pays roughly two to three times as much per completed task to get there.
Ceiling on the hardest tasks
Anthropic's own wording is "comes close to" the frontier — near-frontier, by the vendor's own description.
Still the top of the Claude range, and the Mythos class remains ahead of Opus 5 on cybersecurity evaluations.
Knowledge freshness
Reliable knowledge cutoff May 2026.
Reliable knowledge cutoff January 2026 — four months older.
Caching economics
Cache hits and refreshes at $0.50 per million tokens, and the minimum cacheable prompt drops to 512 tokens.
Cache hits and refreshes at $1 per million tokens, with the usual 1,024-token minimum.
Latency profile
Rated "moderate" on Anthropic's comparative-latency scale.
Rated "slower" — the price of the extra ceiling.
Context window and output limit
1M-token context window (both the default and the maximum, no smaller variant) and 128k maximum output.
1M-token context window and 128k maximum output — identical.
Strength of the evidence
Two days old at the time of writing: the headline numbers are Anthropic's own charts, and BenchLM lists its #1 placement with an "Estimated" evidence label.
Generally available since June 2026, with a public track record; BenchLM lists its #3 placement as "Supported".
Total Score5/ 82/ 81 ties
List API price
The new default
$5 per million input tokens, $25 per million output tokens.
The frontier ceiling
$10 per million input tokens, $50 per million output tokens — exactly double.
Cost per completed task
The new default
On CursorBench 3.2 at max effort it lands within 0.5% of Fable 5's peak score at half the cost per task; on OSWorld 2.0 it beats Fable 5's best result at just over a third of the cost.
The frontier ceiling
Holds the peak score on those same benchmarks, but pays roughly two to three times as much per completed task to get there.
Ceiling on the hardest tasks
The new default
Anthropic's own wording is "comes close to" the frontier — near-frontier, by the vendor's own description.
The frontier ceiling
Still the top of the Claude range, and the Mythos class remains ahead of Opus 5 on cybersecurity evaluations.
Knowledge freshness
The new default
Reliable knowledge cutoff May 2026.
The frontier ceiling
Reliable knowledge cutoff January 2026 — four months older.
Caching economics
The new default
Cache hits and refreshes at $0.50 per million tokens, and the minimum cacheable prompt drops to 512 tokens.
The frontier ceiling
Cache hits and refreshes at $1 per million tokens, with the usual 1,024-token minimum.
Latency profile
The new default
Rated "moderate" on Anthropic's comparative-latency scale.
The frontier ceiling
Rated "slower" — the price of the extra ceiling.
Context window and output limit
The new default
1M-token context window (both the default and the maximum, no smaller variant) and 128k maximum output.
The frontier ceiling
1M-token context window and 128k maximum output — identical.
Strength of the evidence
The new default
Two days old at the time of writing: the headline numbers are Anthropic's own charts, and BenchLM lists its #1 placement with an "Estimated" evidence label.
The frontier ceiling
Generally available since June 2026, with a public track record; BenchLM lists its #3 placement as "Supported".

Key Statistics

Real data from verified industry sources to support your decision.

Claude Opus 5 lists at $5 / $25 per million input and output tokens, Claude Fable 5 at $10 / $50 — exactly double

Claude Platform Docs — Pricing

On CursorBench 3.2 at max effort, Opus 5 scores within 0.5% of Fable 5's peak at half the cost per task

Anthropic — Introducing Claude Opus 5

On OSWorld 2.0, Opus 5 surpasses Fable 5's best result at just over a third of the cost

Anthropic — Introducing Claude Opus 5

On ARC-AGI 3, Opus 5 scores three times as high as the next-best model

Anthropic — Introducing Claude Opus 5

Reliable knowledge cutoff: May 2026 for Opus 5 against January 2026 for Fable 5, with both models on a 1M-token context window

Claude Platform Docs — Models overview

Public leaderboard placement: Opus 5 at #1 carries an "Estimated" evidence label, Fable 5 at #3 a "Supported" one

BenchLM

All statistics come from verified third-party sources. Source, year, and direct link are shown on each metric.

When to Choose Each Option

Clear guidance based on your specific situation and needs.

Choose The new default when...

  • You run high-volume agentic or coding work where the cost per completed task decides the budget, not the top of a leaderboard.
  • Your agents lean on prompt caching: the $0.50 cache-hit rate and the 512-token cache minimum compound across long sessions.
  • Your workload touches anything between January and May 2026 and you would otherwise pay for retrieval to close that gap.
  • You are already on Opus 4.8 — Opus 5 costs the same, so the upgrade is an evaluation question, not a budget request.

Choose The frontier ceiling when...

  • You need the highest available score on the hardest tasks and the last half percent is worth paying double for.
  • Your evaluation policy requires a public track record: Fable 5 has been generally available since June 2026, Opus 5 for days.
  • Your work is security-heavy — the Mythos class remains ahead of Opus 5 on cybersecurity evaluations.
  • Your pipeline is tuned and validated against Fable 5's behaviour and a full re-evaluation would cost more than the price difference saves.

Our Recommendation

Opus 5 is the right default for almost everyone, and the argument is not that it is the smarter model — it is that it removes the reason to pay double. Half the list price, half the cache-hit price, a lower cache minimum, a four-month fresher knowledge cutoff and a faster latency rating are not marginal gains that need a benchmark to justify them; they show up on the invoice and in agent runtime whether or not the leaderboard moves. On Anthropic's own task-level numbers the cost story is stronger still: half the cost per task on CursorBench 3.2 for a score within 0.5% of Fable 5's peak, and a third of the cost on OSWorld 2.0 for a better result. What Fable 5 keeps is narrow but real. It still holds the peak on the benchmark Anthropic used to make the cost argument, the Mythos class remains ahead on cybersecurity evaluations, and it has a public record where Opus 5 has a launch post. That last point deserves more weight than it usually gets: at the time of writing, the numbers that make Opus 5 look dominant are Anthropic's own charts, published two days ago, and the one third-party leaderboard we checked labels its #1 placement "Estimated" while Fable 5's lower placement is "Supported". That is not evidence that the claims are wrong — it is a reason to verify them on your own workload before you re-platform a production pipeline. The practical decision splits cleanly. If you are on Opus 4.8, move: it is the same price and a strictly better model by every published measure, so the only work is re-checking max_tokens now that thinking is on by default. If you are on Fable 5, run one honest evaluation before switching — if Opus 5 holds on your tasks, you halve a recurring bill; if your workload lives in the last half percent, you have just paid for the evidence that Fable 5 is worth its price. And if you were about to pay Fable rates purely for speed, note that Opus 5's Fast mode costs exactly the same $10 / $50, which makes that a direct comparison rather than a tier decision.

Frequently Asked Questions

Common questions about this comparison answered.

On most published evaluations Opus 5 leads, and Anthropic calls it state of the art on coding and knowledge-work benchmarks such as Frontier-Bench and GDPval-AA. But the vendor's own phrasing is that Opus 5 "comes close to" Fable 5's frontier intelligence, and on CursorBench 3.2 the peak score still belongs to Fable 5 — Opus 5 gets within 0.5% of it for half the cost. If you optimise for intelligence per euro, Opus 5 wins clearly. If you optimise for the absolute ceiling, Fable 5 has not been displaced.
Exactly half on list price: $5 / $25 per million input and output tokens against $10 / $50, and cache hits at $0.50 against $1. On Anthropic's own task-level measurements the gap is similar or larger — half the cost per task on CursorBench 3.2, just over a third on OSWorld 2.0. One exception matters: Fast mode for Opus 5, currently a research preview on the Claude API only, is priced at $10 / $50, which is exactly Fable 5's rate.
No. Both run a 1M-token context window and a 128k-token maximum output. On Opus 5 the 1M window is both the default and the maximum, with no smaller variant to choose. The real differences are price, comparative latency (moderate against slower), the knowledge cutoff, and how each model handles thinking and effort.
Three things. Thinking is on by default on Opus 5, and max_tokens is a hard limit on thinking plus response text, so revisit any max_tokens value tuned for a non-thinking setup. The effort ladder — low, medium, high, xhigh, max — is your quality-versus-cost dial, and Anthropic recommends starting at the default of high and adjusting from your own evaluations. And run those evaluations: the published margins are two days old and largely first-party.

Need help deciding?

Book a free 30-minute consultation and we'll help you determine the best approach for your specific project.

Free consultation
No obligation
Response within 24h